Gemini 4 Argon: How Google's New Flagship AI Model Works and What It Does

Google released Gemini 4 Argon to developers, businesses and consumers worldwide on October 13, 2026, according to the company's own announcement on the Google blog. The company describes it as a frontier model aimed at "deep work" — complex, multi-step tasks in coding, finance, legal work, cybersecurity and science — rather than short question-and-answer exchanges.

What Gemini 4 Argon is built to do

According to Google's launch post, Argon combines deep reasoning over multi-step tasks with multimodal understanding of documents, images and spatial environments, plus the ability to act as an autonomous software agent. The model supports what Google calls a 1-million-token "unified" input-output limit, meaning a single exchange can both read and generate very large amounts of text, code or data without losing context.

Google reports that Argon ranks first on the Vals Index, a benchmark the company says measures performance on economically important tasks such as finance, coding, legal work and tax preparation. The company also cites a score of 87% on the Embodied Reasoning (ERQA) benchmark, which tests spatial and physical problem-solving, and says Argon showed a vulnerability-unlock rate of under 1% in bioweapon-related red-teaming tests.

Coding: from Chrome's codebase to GitHub Copilot

Google's announcement names several internal and customer use cases: engineers used Argon to help separate a roughly 6-million-line legacy folder of browser features from within Chromium's larger, roughly 40-million-line codebase, and a Python workflow optimization task reportedly cut session-preparation time by 40%. A quantum-chemistry application is said to have run 2.3 times faster with Argon's help.

On October 13, 2026, GitHub announced that Gemini 4 Argon became generally available inside GitHub Copilot for Pro+, Business and Enterprise plans, selectable from the model picker in Visual Studio Code, Visual Studio, the Copilot CLI, JetBrains IDEs, Xcode and Eclipse, according to the GitHub changelog. GitHub says that in its early testing, Argon solved more coding-benchmark tasks than previous Gemini models, with its largest gains in the Copilot CLI, and that it checks its work by testing edge cases before finishing. Billing for Copilot users runs on usage-based pricing at the provider's list rates.

Robotics and physical reasoning

A separate post on the Google Developers blog focuses on Argon's robotics applications. Google says the model can translate real-world robot camera footage, human point-of-view video or even hand-drawn sketches into physically calibrated simulated environments, a process it calls "Real2Sim." In one demonstration, Argon directed a Boston Dynamics Spot quadruped robot to retrieve and dispose of an object after practicing the task inside a MuJoCo physics simulation.

Google also says Argon can generate synthetic training data inside simulation to improve other robot-control models, reporting that one vision-language-action model's task success rate rose from 56% to 75% after training on Argon-generated data. The model is identified in developer tools as gemini-4-argon and, for robotics work, is available through Google AI Studio and Google Antigravity to Google AI Ultra subscribers, and through the Gemini API's paid tier.

How it compares with rival models

Independent benchmark tracking published by the AI-education site DataCamp, which compiled Google's published scores alongside rival releases, found that Argon leads on knowledge-work and long-context tasks but trails competing models on some pure coding and terminal-based benchmarks.

DataCamp's analysis also noted that Argon's standard API price of $4 per million input tokens and $20 per million output tokens matches Claude Opus 5.5's pricing, while its introductory rate of $2/$10 per million tokens undercuts GPT-6 Astra's listed $10/$50 rate during the promotional period.

Pricing, availability and why Google is pushing Argon now

Per Google's own post, introductory API pricing — $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached inputs — runs through November 30, 2026, after which it rises to $4 and $20 per million tokens respectively. Consumers can reach Argon through the Gemini app and AI Mode in Search if they hold an AI Pro or AI Ultra subscription; the Antigravity coding environment is limited to Ultra subscribers, while scientists can get six months of free high-tier Antigravity access and verified cybersecurity professionals can apply through Google's Fairwind Program.

News agency Reuters, in a report carried by Investing.com, placed the launch in a wider context, saying Google rolled out Argon as it sought to regain ground against rivals OpenAI and Anthropic following months of delays and a leadership shakeup at its DeepMind research lab. The report said Google had previously cancelled a planned Gemini 3.5 Pro release that had been slated for June. Early customers named in the report include Deloitte, LexisNexis, Ambience Healthcare and Isomorphic Labs, and teams using Argon inside Antigravity reportedly increased agentic code submissions by more than 35% while cutting code rollbacks by close to a third.